An AI can nail every single step of a task and still fail the whole job. Why?
Why do long-horizon agents fail when their models can solve individual steps?
This explores why an agent that handles each step correctly can still fall apart over a long run, meaning what breaks when steps are strung together rather than in any single move.
This explores why an agent that handles each step correctly can still fall apart over a long run. The corpus mostly points away from "the model isn't smart enough" and toward failures of process: holding a goal, checking work along the way, and not giving up. The weak link is often the loop around the model, not the model.
The first failure is quitting. Across 17 frontier models on 36 expert-designed optimization tasks, the best predictor of success was persistence in repeated benchmark-edit-incorporate cycles, not how good the first attempt was. Most models stopped early or burned their budget unproductively (What predicts success in ultra-long-horizon agent tasks?). The second is wandering. Reasoning models explore invalid paths and switch away from promising ones too soon, abandoning solutions that were viable. That is a problem of organization, not compute, and a simple penalty on switching thoughts improves accuracy without any retraining (Why do reasoning models abandon promising solution paths?). The third is drift. When several agents cooperate, they flip roles, fall into infinite loops, or wander off-topic, because LLMs have no persistent representation of their goal or role (Why do autonomous LLM agents fail in predictable ways?).
This is easy to miss because it doesn't look like a wrong answer. One study found most long-trace failures were process violations, not incorrect outputs. Checking intermediate states and policy compliance during generation raised task success from 32% to 87% (Where do reasoning agents actually fail during long traces?). A test that scores single steps or final answers can pass a model that reliably derails in the middle.
The fixes come from two directions. One builds structure around the model. Reliable agents move memory, reusable skills, and interaction protocols into a harness, so the model doesn't have to re-solve the same problems each time (Where does agent reliability actually come from?). Episodic memory alone lets an agent keep improving without any weight updates, reaching 87.88% on GAIA validation (Can agents learn continuously from experience without updating weights?). The other direction is training for the long game. Most LLM training treats the model as a one-step generator, while agentic RL treats it as a policy acting over many turns with partial information (How does treating LLMs as multi-step agents change what we can optimize?). That matters because an end-of-task reward can't say which step went wrong. Step-wise expert-similarity rewards give a learning signal even when every attempt fails (Can step-wise expert rewards help small models learn hard reasoning?). Per-turn credit can also come from how much an agent's own belief shifts toward the answer (Can an agent's own beliefs guide credit assignment without critics?). Failed episodes can be kept as abstracted lessons rather than raw transcripts (Should successful and failed episodes be processed differently?).
The notes here don't compare the scaffolding camp and the training camp head to head, so it's unclear which fix matters more. What they agree on is that solving a step and sustaining a run are different skills.
Sources 10 notes
Across 17 frontier models on 36 expert-curated optimization tasks, repeated benchmark-edit-incorporate cycles within a wall-clock budget proved the dominant success predictor. Most models terminated early or burned budget unproductively; Claude Opus 4.6 stood out as persistent.
Reasoning LLMs exhibit two reinforcing failures: wandering (invalid exploration) and underthinking (premature path-switching). Decoding-level interventions like thought-switching penalties improve accuracy without fine-tuning, suggesting viable solutions exist but are abandoned prematurely.
Research identifies role flipping, flake replies, infinite loops, and conversation deviation as LLM-specific failures in multi-agent cooperation. These occur because LLMs lack persistent goal representation and stable role identity.
Reliability for long-trace reasoning comes from checking intermediate states and policy compliance during generation, not from scoring final outputs. Adding intermediate verification raised task success from 32% to 87% because most failures are process violations, not wrong answers.
Research shows reliable LLM agents externalize three cognitive burdens—memory (state persistence), skills (procedural components), and protocols (structured interaction)—into a harness layer rather than relying on model scale alone. The harness unifies these externalities and eliminates the need for the model to solve the same problems repeatedly.
Show all 10 sources
AgentFly formalizes agent learning as a Memory-augmented MDP with three memory modules (case, subtask, tool) that enable credit assignment and policy improvement entirely through memory operations. The approach achieved 87.88% on GAIA validation without modifying LLM parameters.
The Agentic RL survey shows that modeling LLMs as policies in Partially Observable MDPs rather than single-step generators makes memory, planning, and reasoning into RL-optimizable subsystems. This structural reframing explains the recent empirical convergence across memory-based agents, skill learning, and strategy distillation.
Supervised Reinforcement Learning rewards models by measuring alignment with expert actions at each step, providing dense learning signals even when all rollouts fail. This approach bridges the gap between rigid token-by-token imitation (SFT) and sparse outcome-only rewards (RLVR), and works best as a curriculum foundation before outcome-based refinement.
ΔBelief-RL uses log-ratios of sequential probability estimates to assign per-turn credit without critic networks or process reward models. Tested on 20 Questions, smaller models trained this way matched or exceeded prior SOTA and larger baselines while generalizing beyond training.
SkillRL demonstrates that treating successful episodes as concrete demonstrations and failures as abstracted lessons achieves state-of-the-art performance on complex tasks while using substantially less context than uniform approaches. The asymmetry mirrors human expert reasoning and avoids the degradation seen in uniform consolidation methods.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- AgentFly: Fine-tuning LLM Agents without Fine-tuning LLMs
- Useful Memories Become Faulty When Continuously Updated by LLMs
- RLVMR: Reinforcement Learning with Verifiable Meta-Reasoning Rewards for Robust Long-Horizon Agents
- ReasoningBank: Scaling Agent Self-Evolving with Reasoning Memory
- Large Language Model Reasoning Failures
- SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
- LatentSkill: From In-Context Textual Skills to In-Weight Latent Skills for LLM Agents
- Bilevel Coordinated Reflection: A Game-Theoretic Approach to Multi-Agent LLM Systems